Inferring the location of neurons within an artificial network from their activity

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Inferring biological neural network connectivity from activation data is challenging.
• Artificial neural networks offer a tractable model for studying network inference.
• Understanding artificial neural network structure can inform biological network analysis.

Purpose of the Study:

• To investigate the problem of assigning artificial neurons to specific locations within a known network architecture (LeNet image classifier).
• To develop and evaluate a supervised learning method for neuron localization based on activation patterns.
• To identify characteristics of effective datasets for accurate neuron assignment.

Main Methods:

• Utilized a supervised learning approach for neuron localization.
• Derived features from eigenvectors of the neural activation correlation matrix.
• Tested performance using various image datasets with the LeNet architecture.

Main Results:

• The effectiveness of a dataset for neuron localization depends on its ability to fully activate the network and minimize confounding correlations.
• No single image dataset achieved perfect neuron assignment.
• Combining features from multiple image datasets resulted in perfect neuron assignment.

Conclusions:

• Artificial neural networks provide a valuable framework for studying network inference problems.
• Dataset properties significantly impact the accuracy of neuron localization in artificial networks.
• A multi-dataset approach enhances the precision of neuron assignment within artificial neural networks.